Multivariate TS classification

[zeisal]

Dear Ben,
first of all thank you for all the work and effort you put into the hctsa package. It has been really helpful so far!

In the beginning I was dealing with a single sensor to monitor a pressure ts. The classification is mostly a 2-class but can also be a multiple class problem.
Now, I have multiple sensors available that monitor one and the same process. I would like to include possible relations and dependencies between the sampled variables into my analsis. Therefore, I like to ask if you have made any experience yet on how to implement a multivariate ts analysis in a smart way, still using all the beautiful functions within this package.

Thank you in advance and best regards,
Alex

[benfulcher]

Thanks Alex for the kind words and glad you're enjoying hctsa 😄
hctsa is designed for univariate time-series analysis, so there is no native support for the types of multivariate analyses your mention.
You have some options:

1. Append lots of univariate feature matrices together (so that features of each univariate time series are captured). [you may wish to do some dimensionality reduction in this case]
2. Add some features of multivariate dependencies manually.
   Neither of these options are natively supported at this stage, but you could use hctsa simply as a way to compute the feature matrix, that you then analyze yourself externally.

Hope this helps,
Ben

[zeisal]

Thank you Ben for your quick reply.
I thought about the same. Since I used the reduced feature set of only ~220 features I still have an overseeable matrix even with multiple ts.
I also came across the question on why you selected especially those features for the reduced set. Is this from your experience or how did you select them?
Best regards, Alex

[benfulcher]

Hi Alex,
The reduced feature set of 220 is very rough and outdated.
See here for an updated pipeline on this topic (and associated paper).
Best,
Ben